Does adding any noise protect privacy?

No. Protection depends on the mechanism, release pattern, and threat model.

Can categorical values be perturbed?

Yes, through probabilistic replacement rules.

Does zero-mean noise preserve every statistic?

No. It preserves a simple additive value in expectation under the stated model, not all derived analyses.

AI Glossary / Statistics and probability / Intermediate

Random perturbation

Definition

Random perturbation, in statistical disclosure control, deliberately changes data using a random mechanism before release. Numerical values may receive random noise, while categorical values may be replaced according to specified probabilities. The aim is to reduce the risk of revealing confidential information while retaining useful statistical properties. This entry follows the supplied OECD glossary's disclosure-control meaning. The broader expression can also describe random changes in optimization or numerical analysis, which are different applications.

What it means

Released values need not be exact copies of confidential observations to support useful analysis. A carefully designed perturbation mechanism can obscure individual contributions while preserving some patterns at an aggregate level.

The protection and usefulness depend on the entire mechanism, not simply on whether some randomness was added. Noise that is too small, repeated releases that can be combined, or auxiliary information may undermine protection. Conversely, excessive noise can obscure meaningful differences. A perturbed value is therefore a statistical output with known methodological qualifications, not an exact statement about a particular person or business.

How it works

Specify what will be released, who could be harmed by disclosure, and what information an attacker might already possess. Choose whether to perturb underlying records or published summaries. For additive numerical perturbation, a simple form is released value = original value + noise. A mean-zero noise distribution can preserve the original value in expectation, but a particular release still differs from it.

For categorical data, define transition probabilities that govern retention or replacement of each category. Such changes may alter marginal frequencies and associations unless the mechanism and subsequent analysis account for them. Numerical mechanisms can likewise affect variances, correlations, small counts, and consistency between totals.

Evaluate confidentiality risk and statistical utility together. Document constraints, any post-processing, and how uncertainty should be interpreted. Check the combined effect of all releases. Random perturbation is a broad family of techniques; differential privacy is a specific formal guarantee that requires a properly specified and calibrated mechanism. Arbitrary noise addition does not establish that guarantee.

Three practical examples

1. A statistical publisher explores adding noise to aggregate business counts while evaluating how much useful regional information remains.

2. A data custodian considers probabilistic replacement of sensitive categorical responses, then assesses whether associations needed by researchers remain interpretable.

3. An internal analytics team tests perturbed summaries for external sharing and compares errors in large groups with errors in small groups before considering release.

Think of it this way

Think of deliberately blurring a map before sharing it. The blur may hide a precise location while preserving the broad layout, but its effectiveness depends on the scale and on what other maps the recipient possesses. Random perturbation similarly requires a defined protection goal and an assessment of what remains inferable.

Worked example

Consider an original toy mechanism for a confidential count of 50. Add noise E taking values -2, 0, or +2 with probabilities 0.25, 0.50, and 0.25. The possible released counts are 48, 50, and 52.

The expected noise is (-2)(0.25) + 0(0.50) + 2(0.25) = 0, so the expected released count remains 50. The noise variance is 4(0.25) + 0 + 4(0.25) = 2, with standard deviation sqrt(2), approximately 1.414. The expected absolute perturbation is 2(0.25) + 2(0.25) = 1.

Suppose the same confidential count is released 100 times with independent fresh noise from this distribution. Averaging those releases produces noise variance 2/100 = 0.02 and standard deviation approximately 0.1414. Repetition can therefore make the original count much easier to estimate.

This bounded-noise construction is a teaching example, not an approved disclosure-control method or a claim of differential privacy. It demonstrates expectation, variance, and the danger of treating repeated releases separately. It also omits boundary handling for small counts, which would need explicit design and could change bias.

Illustrative case study

In a fictional trade association, analysts plan to publish monthly member-business statistics. They initially propose adding small random changes to each number and assume that this automatically makes release safe.

A disclosure review shows that overlapping tables and repeated versions could reveal much more than any single table suggests. The team assesses the release program as a whole, tests effects on small sectors, and documents how published totals relate to component values. It postpones publication until an appropriate protection method and interpretation guidance are established. No claim of guaranteed anonymity is made merely because the figures have changed.

Common misconceptions and mistakes

Mean-zero noise does not mean a released dataset is unchanged or that every statistic remains unbiased. Nonlinear transformations, clipping, ratios, and selection rules can introduce bias. Independent noise added to cells can also make totals inconsistent unless a reconciliation method is used.

Do not equate perturbation with removal of identifiers, anonymization, encryption, or differential privacy. Each describes a different operation or guarantee. Reusing exact confidential totals alongside noisy subtotals may reveal information. Repeated independent releases can average away noise, while reusing noise has its own design implications. Confidentiality assessment must consider the complete release context rather than a single altered number.

Business use

Random perturbation is relevant when organizations want to share statistical information without publishing exact sensitive values. Its business value depends on maintaining enough accuracy for the intended decisions while meeting applicable protection requirements. Use qualified disclosure-control review, assess small groups and unusual records, and explain limitations to data users. Method suitability and institutional policy must be checked for the specific release rather than inferred from historical examples.

Check your understanding

Independent zero-mean noise has variance 9 in each of 36 releases of the same value. What is the noise standard deviation of their average?

Answer: It is sqrt(9/36) = 0.5. This illustrates why repeated releases must be considered jointly.

Related terms

Statistical Disclosure Control (SDC)
Differential privacy
Random rounding
Statistical disclosure control
noise infusion
confidentiality
data utility
post-randomization method
disclosure risk

Learn next

Study disclosure risk and utility assessment, then compare perturbation, suppression, and coarsening. Learn the formal requirements of differential privacy before interpreting a noise-based mechanism as providing that guarantee.

Differential privacy

Sources

OECD, Glossary of Statistical Terms, Random perturbation and Perturbation-based methods: https://www.oecd.org/content/dam/oecd/en/publications/reports/2008/09/oecd-glossary-of-statistical-terms_g1gh9ad7/9789264055087-en.pdf

The official glossary's disclosure-control definition was read directly. The numerical mechanism and fictional case are original, deliberately simplified illustrations, not recommendations for a production privacy system.

Previous: Random optimization Next: Random policy